AI Disclosure | Derive

This notice is published by Derive Notes Pty Ltd (ACN 696 859 597), 2/290 Boundary Street, Spring Hill QLD 4000, Australia. It sits under the Privacy Policy and the Terms of Service. Questions go to matt@derivenotes.com.

The short version

Derive reads your handwriting using AI, and that is not optional, because reading your handwriting is the product. Some of the extra features built on top of it are optional and can be turned off. AI features are not gated on your consent: they run because you use the product. AI output can be wrong and you should check it.

Where AI is used

Always on. This is how Derive works.

Reading your handwriting. An image of the region you wrote in, plus the recognised text of the lines around it, is sent to an AI provider and comes back as text and maths.
Deciding which lines on the page are relevant. The recognised text of nearby lines is sent so the model knows what you are working on.
Reading a photo or scan you add, so we can pull out the questions and definitions on it.
Checking your working and answering a question about the page. This is the tutor.
Building the search index. Your page text is turned into a numeric representation so that search works.

Optional. Controlled by the Handwriting recognition switch in Settings, AI.

Summaries of a page
Automatic page titles
Practice questions
Diagram generation
Pulling text or code out of an image
Explaining code
The assistant chat
Semantic search over your notes

Only if you set it up yourself.

Connecting Derive to claude.ai or Claude Desktop, which lets Claude search and read your notes.

Code you run. Code you write in the app and choose to run is sent to Judge0, through RapidAPI, to be executed. The execution is limited to 10 seconds of wall time and 128 MB of memory.

What leaves your device

For every AI feature, the request carries page content and nothing that identifies you to the provider. Depending on the feature, that content is:

a rendered image of the region of the page you wrote on,
the recognised text of the lines around it,
the typed text, title and structure of the page,
a photo or scan you added,
the question you typed to the tutor or the assistant,
source code you wrote, and the input you gave it,
on a practice page, the question you are working on.

We never send your name, your email address or your Derive account id to any AI provider.

Who processes your content

What you do	Who receives it
Write on the page and use recognition or the tutor	OpenRouter, and from there Relace, Modal or DeepInfra. If those fail, Google, then OpenAI
Add a photo of a textbook page or a problem set	Google, or OpenRouter
Ask for a summary, a title, practice questions, a diagram, or use the assistant	OpenAI, and Google for search indexing
Search your notes	Google, which turns your search text into a numeric representation
Run code	Judge0, through RapidAPI
Connect Claude	Anthropic

Where each company processes the data, and a link to each company's own terms, is on the Subprocessors page.

What the providers may do with it

We use these providers on their standard commercial API terms, which limit their use of your content to providing the service to us. On the handwriting recognition path we additionally send an instruction with every request refusing any model host that would keep or train on the request. Requests on the other paths, including OpenAI and Google, do not carry a zero retention instruction and run on those providers' standard terms.

We do not control their systems and cannot guarantee their behaviour.

Training our own models

Separately from the providers above, we use content you create to train and evaluate our own handwriting and mathematics recognition models. This is on by default. You can object by emailing matt@derivenotes.com. It is a manual process and there is no switch in the app. Full detail, including what is collected, is in Children and students and the Privacy Policy.

AI output can be wrong

Recognition can misread handwriting, especially dense algebra.
The tutor can mark correct working as wrong, and wrong working as correct.
Summaries, practice questions and diagrams can be inaccurate or incomplete.
Nothing Derive produces is professional, academic, legal, financial or medical advice.
Check anything you rely on, and do not submit AI output as your own work where your school or university's rules do not allow it.

When the tutor marks a step, you can tell it that it is wrong. We record that, and we use it to make the model better.

Automated decisions

Derive uses automated systems to group your strokes into lines, to read them, and to decide whether to mark a step as wrong or to change a verdict it previously gave. Those decisions are about your maths working. They have no legal effect and no similarly significant effect on you, they do not affect your access to Derive, and you can always overrule the tutor.

Marking AI generated content

We do not currently attach machine readable provenance information to content Derive generates for you, such as a generated diagram or a generated summary. The feature that produced it is named on screen when you use it.

What you can turn off

Thing	Can you turn it off	Where
Handwriting recognition and the tutor	No. Derive does not work without it	Not applicable
Summaries, titles, practice questions, diagrams, image extraction, assistant, semantic search	Yes	Settings, AI, "Handwriting recognition"
Usage analytics	Yes	Settings, Account
Screen recording	It is off for everyone, and there is no control in the app that turns it on	Not applicable
Crash and performance diagnostics	No	Not applicable
Use of your work to train our recognition models	By email to matt@derivenotes.com. There is not yet a switch	See "Training our own models" above
The Claude connector	Yes, it is off unless you connect it	claude.ai connector settings